articles
Distilled pieces drawn from the data in the longer Analysis work. One argument each, the numbers that carry it, no filler. RSS.
A small model runs as our junior coding agent on our own hardware. Four mechanisms around it decide what it can accomplish, covering commands it cannot reach, procedures it recalls, a fixed suite that scores it, and a breaker that escalates. None of the parts are novel. The assembly turns the outcome of a request from a probability into a guarantee.
A single layer of a 235-billion-parameter model weighs a gigabyte and a half, more than an entire small dense model. A browser tab holds one now, streaming it in a tensor at a time so the layer never sits whole in memory. Peak memory drops from 1.5 gibibytes to 630 mebibytes, shipped in Unstable Legion v0.1.5 and live.
The House passed permanent daylight saving time on 14 July 2026. Does it stand a real chance, what would it cost or save, and what should an organization actually do about it?
One language model, cut into layer-stages that run in ordinary web browsers and hand work to each other over the Codec wire, with no data center in the loop. The pipeline assembles itself across whoever is online and keeps generating as peers come and go. It is early and experimental, and it runs today at legion.codecai.net.
Every Fourth of July celebrates an independence most of us have quietly traded back, one subscription at a time. What does it actually cost to move your files, your photos, your mail, and your AI into your own house? And why would the technology sold as the ultimate instrument of control turn out to be the shortest road out?
Companies are bolting retrieval systems onto their AI stacks to stop the model reinventing the same code on every request. The cheap version is older than the expensive one: keep a small library of the deterministic things you already solved, and let the model recall a proven tool instead of gambling on a fresh one. A skill is a prompt. A recalled tool is a guarantee.
A computer-science look at the recognitions you cannot be taught, only earn: the patterns and tools whose purpose stays invisible until scale or time forces it. Idempotency, eventual consistency, the index, the interface, and the exact moment each one finally makes sense. Plus why AI is quietly removing the loop that produces that understanding.
AI keeps getting cheaper, the bills keep getting bigger, and the build-out still cannot earn itself back. Why falling token prices squeeze the companies that paid for the data centers instead of saving them.
The AI-infrastructure bet is rational even if AI underdelivers, because of two doors most coverage never connects. The front door is index construction, the machinery now being loosened to load the public into mega-bets like SpaceX. The back door is the documented federal pattern of socializing losses. Private gains, socialized losses, priced into the return math before the bet is placed.
The thing AI makes possible that Agile could only promise: a silent intermediary that sits between every expert in a software organization, translates each one's dialect, shows each only the concerns that touch them, and lets a unit of work clear every reviewer in parallel instead of through a chain of meetings. Plus the part no one else is building, a pipeline that turns juniors into seniors on purpose.
NVIDIA earns $58 billion in a quarter selling the chips that fund the cloud, and now sells the laptop chips that obsolete it. The one company guaranteed to win the AI buildout is the one underwriting both sides of the trade.
Every time anyone uses AI, the cloud converts the model's native numbers into text, wraps them in web packaging, and ships the bundle to a reader who is usually another machine that throws the packaging away. It costs the industry $400 million a year and rising, strains the same grids and water tables the data-center boom already strains, and prices billions of people out of AI entirely. The fix is small.
Companies are spending heavily to give AI architectural awareness: vector embeddings, semantic search, retrieval pipelines. Good engineering already solved most of it. Put the context next to the code, and let your rules point upstream, and the model honors your architecture with no retrieval system at all. A consolidation of the inline-documentation and defined-contract pieces.
We replaced cursive with typing, but never questioned the 1870s QWERTY layout we teach on. Now the machine on the other side of the keyboard is AI, and the throughput and precision with which we express ourselves to it will define a professional ceiling. The case for taking stenography seriously.
AI does not solve software's oldest problem, the degradation of intent as it passes from stakeholder to product manager to developer. It inherits it, and executes the inherited misunderstanding at a speed and scale no human team ever could. The first in a series on communicating with AI as a practitioner.
Ken Thompson proved in 1984 that a compiler could hide a backdoor from anyone reading the source. AI models now hold that position. Distillation carries whatever they inherit into the next generation. Nobody can read a weight.